Computational Training for the Next Generation of Neuroscientists
Total Page：16
File Type：pdf, Size：1020Kb
Computational training for the next generation of neuroscientists The MIT Faculty has made this article openly available. Please share how this access benefits you. Your story matters. Citation Goldman, Mark S, and Michale S Fee. “Computational Training for the Next Generation of Neuroscientists.” Current Opinion in Neurobiology 46 (October 2017): 25–30 © 2017 Elsevier Ltd As Published https://doi.org/10.1016/j.conb.2017.06.007 Publisher Elsevier Version Author's final manuscript Citable link http://hdl.handle.net/1721.1/118398 Terms of Use Creative Commons Attribution-NonCommercial-NoDerivs License Detailed Terms http://creativecommons.org/licenses/by-nc-nd/4.0/ Highlights: • Reports results of a survey on training needs in computational neuroscience • More quantitative training is needed for students from life science backgrounds • Experience with real biological data is needed for non-life science students • A well-organized, centralized repository is needed to host training resources • Cultural barriers are holding back widespread computational neuroscience training Title: Computational Neuroscience Training for the Next Generation of Neuroscientists Authors names & affiliations: Mark S. Goldman1, Michale S. Fee2 1Center for Neuroscience, Department of Neurobiology, Physiology, and Behavior, and Department of Ophthalmology and Vision Science, University of California, Davis, Davis, CA 95618, USA, [email protected] 2McGovern Institute and Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139, USA, [email protected] Corresponding author: Mark S. Goldman, [email protected] 1 Abstract Neuroscience research has become increasingly reliant upon quantitative and computational data analysis and modeling techniques. However, the vast majority of neuroscientists are still trained within the traditional biology curriculum, in which computational and quantitative approaches beyond elementary statistics may be given little emphasis. Here we provide the results of an informal poll of computational and other neuroscientists that sought to identify critical needs, areas for improvement, and educational resources for computational neuroscience training. Motivated by this survey, we suggest steps to facilitate quantitative and computational training for future neuroscientists. Introduction In 1952, in the Journal of Physiology, Hodgkin and Huxley published their famous series of papers describing the biophysical basis of the action potential and laying out the mathematical framework underpinning much of modern cellular neurophysiology [1-5]. Today, sixty-five years later, the qualitative basis of the action potential is widely taught in nearly every introductory neuroscience class. However, the mathematical modeling that was seminal to their work is rarely taught and, because of lack of quantitative training, may be daunting to most biology students, hindering their deeper understanding of the associated concepts. Computational and theoretical approaches shape and inform nearly every level of analysis in neuroscience. These include biophysical and biochemical characterizations of receptor and signaling proteins [6], conductance-based models of single neuron voltage dynamics [7,8], neural network models of circuit dynamics [9,10] and plasticity [11], and statistical approaches to cognition, reasoning, and behavior [12-14]. Computational neuroscience also provides many of the core data analysis techniques used throughout neuroscience, including bioinformatic analyses underlying genome-wide screens [15-17]; statistical analyses of electrophysiological, optical, and non-invasive functional imaging data [18-20]; and signal-processing algorithms underlying brain-machine interfaces and neural prosthetics [21,22]. More theoretically, computational neuroscience provides the intellectual framework within which many of the brain’s computations are now described. Hodgkin and Huxley’s [1] and Rall’s [23,24] frameworks for describing single-neuron computation are classics. Sensory coding studies have been guided by principles of efficient coding [25] and 2 information theory [26] and, more recently, by insights from deep networks [27]. Attractor dynamics provide a conceptual framework for describing memory networks [28,29]. Signal detection theory provides a foundation for studies of decision-making [30-32]. Hebbian and homeostatic learning rules provide a conceptual underpinning for synaptic plasticity studies [33], while reinforcement learning underlies descriptions of neuronal and behavioral data ranging from dopamine signaling [34] to motor control [35,36]. Despite these significant accomplishments, the need for quantitative and computational approaches is growing rapidly. Whereas past studies were considered cutting edge if they recorded from two neurons at a time, recent revolutions in recording technology now allow for simultaneous measurements of the activity of hundreds or thousands of neurons in a single brain area, or even throughout the entire brain of behaving animals [37]. Recent advances in automated electron microscopic imaging of brain tissue will allow large scale neural circuit reconstruction at single-synapse resolution [38]. At the same time, advances in molecular genetics, cell biology, functional imaging, and bioengineering now make it possible to gather data sets that explore a single system, or a single disease, in depth at the molecular, cellular, network, and behavioral levels. These advances will require new methods for the analysis of massive data sets and new theories and models to connect such measurements to underlying computational principles. Neuroscience training must impart future neuroscientists with the core quantitative and computational skills necessary to keep up with these experimental advances, as emphasized by a number of national reports focusing on the future of neuroscience [39-41] and general biology education [42-46]. These skills include not only the ability to perform sophisticated statistical analyses, but also the ability to interpret and build quantitative models, design experiments to test new models and theories, and form collaborations with interdisciplinary teams. Imparting this knowledge presents a significant challenge to neuroscience departments and programs. Survey To develop a more complete picture of the challenges and opportunities facing computational neuroscience education, we conducted an informal poll of a range of leaders in computational neuroscience training, from textbook authors to course directors, program officers, and faculty representing different subfields of computational neuroscience from cellular biophysics to cognitive neuroscience (Supplementary material 1). Our survey asked respondents 3 to give their opinions on three topics: (1) Necessary curricular training for general neuroscience and computational neuroscience-focused students (Table 1), (2) Barriers to training in computational neuroscience (Table 2), and (3) Suggestions for improvements to computational neuroscience training (Table 3). In addition, we used the survey to gather a list of computational neuroscience training resources available to the general community (Supplementary material 2). Below, we summarize the key themes that emerged from the survey responses. We note that the poll consisted of open-ended rather than multiple choice questions. This led to many rich and insightful comments. However, for the tabulation of requisite training topics (Table 1), this format led to some ambiguities in interpretation; namely, it was sometimes unclear, when a respondent failed to mention a particular subject area, whether it was viewed as already standard in most neuroscience program curricula, viewed as unnecessary, or simply overlooked. Many responses also did not clearly differentiate undergraduate and graduate training needs, so we merged these categories in our analysis. Despite these ambiguities, several recurring themes emerged across the set of responses, and we focus our discussion around these. Theme 1: More quantitative training is needed for students from life science backgrounds. The most common refrain from both theorists and experimentalists was that students from life science backgrounds lacked training in quantitative approaches, programming, and algorithmic thinking (Table 2). For general neuroscience students, the most commonly emphasized needs were for further coursework and training in statistics and data analysis, mathematics, and computer programming or computer science. Also emphasized was the need for coursework in computational neuroscience or other biological modeling. Within the category of statistics and data analysis, many respondents explicitly distinguished “data analysis” from statistics per se, emphasizing the need for students to perform hands-on work with real data sets. Within the mathematics curriculum, probability theory and linear algebra were most commonly cited as important subjects. Interestingly, the training needs identified by experimental neuroscientists and theoretical neuroscientists were highly consistent (Table 1). Several respondents emphasized the need, at both the undergraduate and graduate levels, for quantitative classes tailored to students from life science backgrounds. Finally, many respondents who recommended quantitative coursework beyond calculus and introductory statistics emphasized the importance of beginning this training at the undergraduate level. 4 For students planning to work in computational